Heterogeneous task allocation method based on prediction equilibrium ratio and resource constraint

Abstract

The invention discloses a heterogeneous task allocation method based on a prediction balance ratio and resource constraint, and the method comprises the steps: firstly obtaining a load prediction result of a node load in a mobile edge computing MEC cluster through a deep learning algorithm; and secondly, based on a load prediction result, migrating the load through an early warning mechanism of a load migration strategy LMSPER algorithm based on a prediction balance ratio. Finally, MEC cluster heterogeneous task queue data are obtained, task unloading decision is made based on the task queue data and the MEC cluster state, and task unloading distribution is completed. Good adaptability and stability are shown in different environments, stable running of the cluster can be guaranteed, and the task unloading success rate can be guaranteed under the conditions that tasks are complex and cluster resources are in shortage.

Description

Technical Field

[0001] The present invention belongs to the field of edge computing, and in particular relates to a heterogeneous task allocation method based on predicted balance ratio and resource constraints. Background Art

[0002] In the field of Mobile Edge Computing (MEC), task allocation strategy is a key research topic to ensure user experience and system stability. An effective task allocation strategy can reduce the response delay of key tasks and improve user experience by reasonably allocating various tasks. At the same time, it can ensure the successful offloading rate of tasks, improve the processing capacity of the entire cluster, reduce the load of local MEC nodes, and maintain the stability of the system.

[0003] Existing research work has focused on developing different task allocation strategies to cope with sudden local pressure increases and load migration problems in complex heterogeneous environments. These strategies attempt to achieve load migration by monitoring node load status and issuing early warnings, as well as considering task change patterns and potential form transformations. However, existing methods often face two major problems: first, in a highly heterogeneous MEC environment, the differences in performance and resource requirements between different nodes and tasks lead to the complexity of task offloading; second, the mobility and instability of the load source lead to load migration and burst problems. Summary of the invention

[0004] In response to the above problems, the present invention proposes a heterogeneous task allocation method based on predicted balance ratio and resource constraints, which is a comprehensive, two-stage heterogeneous task allocation strategy based on predicted balance ratio and resource constraints. This strategy not only addresses the problem of local task instability caused by load bursts, but also optimizes task offloading and resource management in complex heterogeneous and task mode changes, aiming to improve the overall performance and stability in the MEC environment.

[0005] A heterogeneous task allocation method based on predicted balance ratio and resource constraints, comprising:

[0006] S1. Obtain the load prediction results of the node load in the MEC cluster through deep learning algorithms.

[0007] S2. Based on the load prediction results, the load is migrated through the early warning mechanism of the LMSPER algorithm, a load migration strategy based on the predicted balancing ratio, to prevent the overload of a single node and maintain the stable operation of the system.

[0008] S3, obtain the MEC cluster heterogeneous task queue data, and make task offloading decisions based on the task queue data and MEC cluster status to complete task offloading allocation. The EMC cluster status consists of the node status uploaded in real time by the nodes in each MEC cluster, the load prediction results of the node load in the EMC cluster, and the past node status and load data.

[0009] The task offloading decision ensures that the migrated tasks can be effectively executed on the new nodes through the efficient task offloading allocation of the heterogeneous task offloading HTO algorithm based on resource constraints and meta-reinforcement learning, further optimizing the system performance. The collaborative and complementary characteristics of the two-stage strategy ensure that the mobile edge computing system can better adapt to the dynamically changing load environment, improve the system's resource utilization and task processing efficiency, and ensure stable operation under high load and heterogeneous environments.

[0010] Furthermore, for S2, stage 1: LMSPER algorithm. This algorithm uses the load prediction framework to perform load early warning monitoring and predict the future load conditions of nodes in the MEC cluster to achieve timely load migration. The algorithm includes the following decision-making process:

[0011] S2.1. Load warning based on load prediction results: Based on the obtained load prediction results, the load prediction framework is used to perform load warning monitoring. When the load warning monitors that the workload of a certain node reaches the threshold of 80%, the load migration mechanism is triggered, and the load balancing ratio of each node and the corresponding task is calculated to obtain the load balancing ratio set of the nodes to be processed. The load balancing ratio set contains the load balancing ratio matrix of each node.

[0012] S2.2. Task migration based on balancing ratio results: The load balancing ratio matrix of each node contains the balancing ratio of each task. The balancing ratios of tasks in each node are sorted from small to large, and load migration is performed in the sorted order until the load prediction result of the node is lower than the load threshold.

[0013] Furthermore, for S3, stage 2: HTO algorithm. Through deep reinforcement learning, task features and cluster status are combined to make detailed task offloading decisions. The training of the algorithm includes the following process:

[0014] S4.1, Task sampling: Sample tasks and paths through the task sampler. First, obtain tasks from the set of tasks to be processed and record the resource requirements for these tasks. Then, the task sampler performs path sampling on the sampled tasks, that is, determines the possible execution path and unloading method of the task based on the resource requirements for the task and the current state of the cluster. Finally, the task sampling processor is used to format the sampled task data.

[0015] S4.2, Internal strategy update: Based on the formatted task data, the proximal strategy optimization PPO algorithm is used to optimize the execution path and unloading method, so that the strategy can achieve better performance on the same or similar tasks. This process is a direct reflection of PPO improving the strategy to obtain higher rewards.

[0016] S4.3. External strategy update: The external strategy update process is usually slow. Meta-reinforcement learning uses the data and experience accumulated on multiple tasks to improve the learning strategy of task offloading, ultimately enabling it to better adapt to new tasks.

[0017] Beneficial effects of the present invention:

[0018] The present invention proposes a heterogeneous task offloading strategy based on predicted balance ratio and resource constraints, which is divided into two parts. The first part is a load migration strategy based on predicted balance ratio, which calculates the alarm task balance ratio based on the load prediction result and performs migration. The superiority of the heterogeneous task offloading strategy based on predicted balance ratio and resource constraints over K8S and SWARM strategies is proved from the two aspects of task delay and cluster resource utilization. Experiments show that the LMSPER algorithm can show good adaptability and stability in different environments, and can have the advantages of computing resource optimization. At the same time, when facing the problem of load explosion, it can ensure the stable operation of the cluster and ensure the resource balance of the cluster. The second part is a heterogeneous task offloading algorithm HTO based on resource constraints and meta-reinforcement learning. The algorithm is designed as two internal and external strategies. The internal strategy and the external strategy work simultaneously, so that the meta-reinforcement learning model can quickly adapt to new tasks while mastering a task. The internal strategy is responsible for quickly adapting to the details of specific tasks, while the external strategy optimizes the learning process of the model from a more macro perspective and improves the learning efficiency across tasks. In addition, this paper conducted a large number of experiments from four dimensions: task difficulty, total number of tasks, average task delay, and task offloading success rate. It was proved in multiple dimensions that the HTO algorithm minimizes task delay when cluster resources are abundant, and can also ensure the task offloading success rate when the tasks are complex and cluster resources are scarce. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The overall architecture of heterogeneous task offloading strategy based on predicted balance ratio and resource constraints;

[0020] Figure 2 The change of average task delay with the change of task density under different migration strategies;

[0021] Figure 3 The average resource utilization of the cluster changes over time under different migration strategies. DETAILED DESCRIPTION

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings in the implementation examples of the present invention. The implementation described is only a part of the embodiments of the present invention and cannot cover all implementation examples. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0023] The heterogeneous task allocation method based on predicted balance ratio and resource constraints specifically includes the following steps:

[0024] S1. Obtain the load prediction results of the node load in the MEC cluster through deep learning algorithms.

[0025] S2. Based on the load prediction results, the load is migrated through the early warning mechanism of the LMSPER algorithm, a load migration strategy based on the predicted balancing ratio, to prevent the overload of a single node and maintain the stable operation of the system.

[0026] S3, obtain the heterogeneous task queue data of the MEC cluster, and make task offloading decisions based on the task queue data and the MEC cluster status to complete the task offloading allocation. The EMC cluster status consists of the node status uploaded in real time by the nodes in each MEC cluster, the load prediction results of the node load in the EMC cluster, and the past node status and load data. The task offloading decision is based on the efficient task offloading allocation of the heterogeneous task offloading HTO algorithm based on resource constraints and meta-reinforcement learning.

[0027] 1. Load Migration Strategy Based on Predicted Balancing Ratio LMSPER Algorithm

[0028] In the instruction manual Figure 1 The structure diagram of the LMSPER algorithm is shown in Figure 1. The strategy is divided into two parts: one is the load warning strategy based on spatiotemporal prediction, and the other is the task migration strategy based on the balance ratio. According to the load prediction results, it is judged whether some nodes will exceed the predetermined threshold in the future. When a MEC node exceeds its corresponding threshold, the load migration strategy will be started accordingly, and the balance ratio of all nodes corresponding to each load that needs to be migrated will be calculated. The strategy with the least impact on the cluster is selected to migrate the workload to other MEC nodes.

[0029] First, define the mathematical expression of the load prediction result as:

[0030] P=P1,P2,…,P n

[0031] Ρ represents the current set of prediction results, where P i represents the future load of the i-th node in this round of prediction, P i The mathematical expression is:

[0032] P i =(ι1,ι2,…,ι t )

[0033] Assume P i A total of t time steps are included in the forecast, that is, the forecast window is t, ι j It indicates the load condition of a node at time step j in this round of prediction.

[0034] For each node, the present invention records the five most recent predicted loads d, and then calculates the deviation D each time i , where i represents different rounds. The set Δ is used to represent the load prediction of each node, and the mathematical expression is:

[0035] Δ=D1,D2,…,D n

[0036] D=d1,d2,…,d5

[0037] Set a weight Z for each d, and the sum of the weights is 1. And set the prediction of the more recent round to be more important, and then calculate the weighted average deviation D of the node based on the weight avg , the mathematical expression for calculating the weighted average deviation is:

[0038] D avg =Z1·d1+Z2·d2+Z3·d3+Z4·d4+Z5·d5

[0039] Use the weighted average deviation to correct the current round of forecasts. The mathematical expression is:

[0040] P adjusted =P+D avg

[0041] The present invention hopes that the resource occupancy of the node does not exceed 80% of the node resources, so the mathematical expression for screening out the node set exceeding the threshold is:

[0042] Node=N1,N2,…,N k

[0043] Where N i Indicates that the load of the i-th node will exceed the threshold in the next five minutes. Assuming that the task set on these nodes is Task, the mathematical expression for calculating the balance ratio of each workload on these nodes is:

[0044] Task=T1,T2,…,T k

[0045] T=(t1,t2,…,t m )

[0046] Where T i represents the workload set of the i-th node, and t1 represents the resource demand vector of the task. The remaining resource matrix W is used to represent the remaining resources of the remaining nodes in the edge system. The corresponding mathematical expression is:

[0047] W=(ω1,ω2,…,ω m )

[0048]

[0049] where ω j Represents a node ∈ j The remaining resources Represents a node ∈ j Remaining Resources remaining amount.

[0050] BR=B1,B2,…,B k

[0051]

[0052] BR represents the workload balance ratio set of these nodes, and B represents the balance ratio matrix b of all tasks in the node migrated to each node ij Indicates the balance ratio of the migration of the i-th task on the node to the j-th node. Balance ratio b ij The calculation expression is:

[0053]

[0054] where b ij represents the balance ratio of the i-th task on the node to be migrated to the j-th node, represents the weight of the kth resource of the jth node, represents the expected load of the jth node on the kth resource at a certain moment in the future, represents the resource demand vector of the kth resource when the i-th task is migrated to the j-th node, The calculation expression of is as follows, which calculates the importance of different resources according to the occupancy rate of different resources in the future.

[0055]

[0056] So far, the set of load balancing ratios of the nodes to be processed is obtained. Next, the migration strategy with the least impact will be selected according to the matrix for migration until the expected workload is lower than the load threshold. The load migration algorithm based on load balancing ratio is shown below.

[0057] 1) First, the algorithm evaluates the difference between the predicted value and the actual value by calculating the prediction deviation for each node. Specifically, the prediction deviation is calculated by comparing the difference between the actual load and the predicted load for each node in the past few predictions. These deviations are used to adjust the predicted load value in the future time window. The adjustment method is to apply the weighted average of the deviations to the original prediction value to obtain a more accurate load prediction. This process ensures that the prediction model can better adapt to actual load changes and reduce errors.

[0058] 2) The algorithm uses the adjusted predicted load value to detect whether each node will exceed the predetermined load threshold in the future time window. If the predicted load of a node exceeds the set threshold (such as 80%), the node will be marked as a warning node and a corresponding node warning vector will be generated. This vector is used to indicate which nodes in the system may need to migrate load to prevent overload.

[0059] 3) The algorithm calculates the load balancing ratio of each task after migrating to other nodes for all warning nodes. Specifically, the algorithm evaluates the load balancing of each node after offloading the tasks in the warning node to other nodes. Then, the algorithm sorts the tasks in the warning node according to the degree of influence of the balance ratio, and prioritizes the tasks with the least impact on load balancing for migration. The task migration process will continue until the predicted load of the warning node drops below the threshold, ensuring that the system remains stable in future load fluctuations.

[0060] 2. Heterogeneous Task Offloading (HTO) Algorithm Based on Resource Constraints and Meta-Reinforcement Learning

[0061] The present invention designs a simulation system for environment interaction for meta-reinforcement learning, in which the execution of tasks by nodes is subject to resource constraints and is also affected by the order of the task graph. The simulation system is now defined as follows: In a certain mobile edge topology system E = {∈1,∈2,...,∈ n}, the set of heterogeneous tasks submitted at a certain time is defined as Γ={τ1,τ2,...,τ m}, where the resource requirements, task size, and task type of each task are different. The resource set of the entire mobile edge system is For the task τ i , define the resource feature vector γ i Indicates the resource demand of the task, and the resource demand matrix of the entire task set is represented by Y. i The definition expressions of and Y are:

[0062] Y=(γ1,γ2,…,γ n )

[0063] γ i=(c i1 ,c i2 ,…,c ik )

[0064] where c ij Represents the task τ i Resources to be applied for j demand.

[0065] The mathematical expression of the unloading decision of the external strategy is defined as:

[0066]

[0067] where ρ ij is a Boolean value, ρ ij If True, it means that the task τ i is offloaded to node ∈ j Go up and execute (including local operation), if ρ ij is False, it means that the task τ i Not at node ∈ j Execute on. From this we can see that the mathematical constraints of the Q matrix are:

[0068]

[0069] The execution of tasks is subject to resource requirements, and hard resource requirements must be met before they can be executed on a node. System resources θ are divided into and Two types, including Represents elastic resources, represents hard resources, so the set of heterogeneous tasks submitted at a certain moment can be expressed as:

[0070] Γ=Γ s ∪Γ h

[0071] For the task τ i , it can be scheduled to a node ∈ j The necessary and sufficient condition is that the node ∈ j After completing all task migrations in this round, the remaining hard resources need to satisfy the task τ i Therefore, only if the task τ is satisfied i Only tasks with sufficient remaining resources can be scheduled to nodes ∈ j , and at this time ρ ij It can be True, that is, 1. For a given decision matrix Q, the decision matrix Q can only be correctly executed when the remaining hard resources of the node meet the elastic resource requirements of the task, otherwise the execution fails.

[0072] For the execution reward of a task i in each task graph, the mathematical expression defining the reward function is:

[0073]

[0074] For a task graph, if one of the tasks fails to execute, all subsequent tasks that depend on this task will be considered failed, and the reward value will be set to -500. If a task is executed successfully, the execution cost will be normalized and then negative.

[0075] Reward in Expression ki represents the reward value of the i-th task in the k-th task graph submitted to the environment in this round, cost ki Indicates the total execution cost of this task, cost kmin Indicates the minimum execution cost among all nodes in the current environment, cost kmin Indicates the maximum execution cost among all nodes in the current environment.

[0076] The calculation method of task execution cost is as follows: If executed locally, only the execution cost is calculated. If uploaded to MEC for execution, the data transmission time needs to be calculated, including upload and download. If the task is offloaded to a directly connected MEC for execution, the uplink and downlink time is calculated once; if offloaded to a non-directly connected MEC, it is assumed that one forwarding is required, so the uplink and downlink time is calculated twice. The execution time is affected by the task configuration file, including CPU cycles, data size, position in the DAG topology, transmission size, wireless transmission rate and MEC status. The mathematical expression is:

[0077] cost=cost ki |A k (T,E),Q ki

[0078] Among them A k (T,E) consists of a series of tasks, Q ki It represents the offloading decision of the i-th task in the k-th task graph. The execution order of tasks is sorted according to the dependencies in the task graph.

[0079] The meta-reinforcement learning model algorithm is designed in two parts. The first part is the internal strategy, which is mainly responsible for task-level learning. It aims to quickly adapt and learn a set of similar tasks during the meta-reinforcement learning process. The goal of this strategy is to maximize the reward in the environment of a given task. The update of the internal strategy is based on a small amount of sample data obtained by sampling, so that the Seq2Seq network can quickly adjust its behavior to adapt to new or slightly changed tasks, thereby adapting to the ever-changing MEC environment. The second part is the external strategy, which focuses on learning at the meta level, that is, using experience across multiple tasks to improve the model's learning strategy so that it can better adapt to new tasks. This design of internal and external strategies is designed to enable the model to quickly master a task while responding to changes in tasks at the same time, that is, it can adapt to load migration. The internal strategy aims to master the details of specific tasks so as to allocate tasks, while the external strategy optimizes the model's learning process from a more macro perspective, thereby improving the network's learning of patterns and knowledge across tasks, thereby improving learning efficiency.

[0080] The algorithm training process is divided into two stages: task sampling in the first stage and strategy updating in the second stage. First, the task and path are sampled by the task sampler, and then the task sampling processor is used to format the sampled task data so that it can be received by the Seq2Seq network (sequence prediction network). Then enter the second stage, which is divided into two steps: the first step is to update the internal strategy, and the second step is to update the external strategy. For the update of the internal strategy, this paper uses the PPO strategy for optimization. PPO introduces some constraints in the strategy update to avoid making too large a strategy change in a single update, thereby improving the stability and efficiency of learning. In the present invention, PPO is used as an algorithm for updating the strategy. Its main purpose is to be used for internal strategy update. Using the data sampled from the environment, the PPO algorithm optimizes the parameters of the strategy so that the strategy can obtain better performance on the same or similar tasks. This process is a direct reflection of PPO's improvement of the strategy to obtain higher rewards, and it is also where PPO plays a role as an exploration strategy, that is, to find the optimal solution by exploring different behavioral strategies. For external strategy updates, this process is usually slow, and more data and experience need to be accumulated on multiple tasks.

[0081] Experimental simulation

[0082] 1. Experiment on load migration strategy based on predicted balancing ratio

[0083] Experimental environment:

[0084] In order to enable the simulation experiment to better evaluate the effectiveness of the migration strategy, the present invention selects four real heterogeneous nodes to form an edge cluster during the experimental stage. Table 1 shows the hardware parameter information of the nodes in the experiment. Table 2 shows the task information used.

[0085] Table 1 Heterogeneous node hardware information

[0086]

[0087] Table 2 Task list

[0088]

[0089] Experimental evaluation:

[0090] The performance of the algorithm is evaluated using indicators such as task latency (ms) and resource utilization. The total number of tasks in this section of the experiment remains unchanged, and the order and time nodes of the tasks remain unchanged. The lower the task latency, the better the performance of the migration strategy, and the lower the cluster resource utilization, the higher the energy efficiency of the migration strategy.

[0091] As the instruction manual Figure 2 As shown in the figure, by comparing the three migration strategies of K8S, SWARM and LMSPER, the average task latency changes with the task density under the premise of unchanged total number of tasks, task occurrence order and task occurrence node. Detailed analysis shows that LMSPER exhibits the best performance under all task density conditions, especially under high task density, its stable performance and low latency growth rate indicate its efficient task processing capability. In contrast, although the latency of K8S grows more steadily, its performance is not better than LMSPER in high-density task processing. The SWARM strategy may show good performance in the initial and medium task density, but its latency grows rapidly under high task density, showing a performance bottleneck. Therefore, considering the performance stability and processing capability under high load, LMSPER is a migration strategy that is more suitable for various task density environments, especially when high-load tasks need to be processed, its excellent performance can provide higher efficiency and lower latency.

[0092] As the instruction manual Figure 3As shown in the figure, we can see that, under the same total task setting, for low load conditions, LMSPER can optimize cluster resource utilization and complete tasks faster. Compared with K8S and SWARM strategies, it can quickly complete load tasks with lower cluster resource utilization. This is because LMSPER can sense the computing power of different nodes and predict the load status in the future, so that the load of all nodes in the cluster is more balanced, so as to more reasonably allocate execution nodes for tasks. For high load conditions, that is, when the load suddenly increases, such as when STEP is 1000, LMSPER also has advantages over K8S and SWARM strategies. It can be seen that with the increase of task density, the cluster resource utilization of K8S and SWARM increases very quickly, while the increase of LMSPER is smaller, which is also due to the prediction ability of LMSPER strategy. When there is a load burst, LMSPER can make a good prediction and avoid risks, so that the resource utilization of the entire cluster is maintained at a low level, while the cluster resource utilization of K8S and SWARM strategies will rise sharply, which will reduce the reliability of the cluster. Therefore, under the same cluster setting, LMSPER has the ability to cope with load bursts in complex environments. Moreover, when facing different load pressures, LMSPER can effectively optimize the processing and show strong adaptability and stability.

[0093] 2. Experiment and Evaluation of Heterogeneous Task Offloading Strategies

[0094] The heterogeneous task offloading strategy experiments are designed to test the performance of the algorithm. The following series of experiments compare the performance of the HTO strategy with the global GD strategy in various environment settings under different task complexities. The task volume is divided into three levels: low, medium, and high. The performance is measured by two indicators: the average task delay (the first digit) and the task offloading failure rate.

[0095] Table 3 Task delay and task offloading failure rate in different environments and different task amounts when the task complexity is 20

[0096]

[0097]

[0098] As shown in Table 3, under the background of DAG complexity of 20, global GD shows relatively consistent task latency in different environment settings. However, its latency increases as the number of mobile devices doubles. This indicates that the task processing performance of the global GD algorithm decreases slightly under high load conditions. Compared with the global GD algorithm, the HTO algorithm shows lower task latency, especially when the number of mobile devices increases. This shows the efficiency of the HTO algorithm in offloading tasks and its superior strategy selection. Considering the task failure rate, the failure rate of the global GD algorithm increases with the increase of the number of mobile devices. This implies that the algorithm has limited ability to handle the increased task volume and that the increase in the number of tasks makes the offloading decision more complicated. The failure rate of the HTO algorithm is almost 0% in all scenarios, indicating that its strategy can effectively solve the task offloading problem of complex heterogeneous clusters. Considering the impact of the task volume, for the global GD algorithm, a larger task volume is expected to lead to a slight increase in latency. It is noteworthy that the offloading failure rate also increases with the increase of the task volume, which implies that the efficiency of the algorithm decreases when the load increases. For the HTO algorithm, a larger task volume leads to a smaller average delay, indicating that the algorithm can improve performance under high pressure and thus reduce the overall delay of the system. In summary, when the task complexity is 20, that is, the task is relatively simple, the HTO algorithm provides better performance in maintaining low task delay and offloading failure rate compared to the global GD algorithm.

[0099] Table 4 Task delay and task offloading failure rate in different environments and different task amounts when the task complexity is 30

[0100]

[0101]

[0102] It can be analyzed from Table 4 that the HTO algorithm shows significant advantages over the global GD algorithm in terms of task delay and task offloading failure rate when the task complexity is 30. Specifically, the average task delay of the HTO algorithm in low, medium and high environment settings is approximately 429.25 milliseconds, 546.25 milliseconds and 666 milliseconds, respectively, which is much lower than the 821.25 milliseconds, 1023.25 milliseconds and 1070.25 milliseconds of the global GD algorithm. In addition, the average offloading failure rate of the HTO algorithm is almost zero in all environment settings, with the highest being only 0.20%, while the failure rate of the global GD algorithm is approximately 14.27%, 13.60% and 16.18%, respectively. This shows that the HTO algorithm can not only maintain a low task delay when processing different task amounts, but also significantly reduce the failure rate of task offloading. As the task amount increases, although the task delay of the two algorithms increases, the HTO algorithm grows more slowly and can maintain a low offloading failure rate, proving its high efficiency and reliability in high-load processing. In summary, when the task complexity is 30, that is, the task complexity is general, the HTO algorithm provides better performance in maintaining low task latency and offloading failure rate compared to the global GD algorithm. Especially in the case of high task volume, the HTO algorithm shows more outstanding processing capability and stability.

[0103] Table 5 Task delay and task offloading failure rate in different environments and different task amounts when the task complexity is 40

[0104]

[0105] It can be analyzed from Table 5 that when the task complexity is 40, that is, when the task is relatively complex, the HTO algorithm still shows significant advantages over the global GD algorithm. When the computing resources are relatively tight, HTO can reasonably allocate tasks so that the task failure rate is at a low level. In contrast, the global GD algorithm cannot consider the task unloading success rate. As the computing resources grow, the HTO algorithm can allocate tasks well in the case of high-density tasks so that the task failure rate is 0.0%. This also verifies that the HTO algorithm can adapt well to environmental changes, and as the task density increases, the task delay of HTO is always controlled at a good level. When the cluster computing resources are tight, for example, when the number of mobiles is 100 and the number of mecs is 50, the HTO algorithm can sacrifice a certain task delay to ensure the cluster task unloading success rate, which is crucial for MEC clusters. In summary, the HTO algorithm can still reasonably allocate tasks when the tasks are relatively complex, and make a reasonable balance between task delay and task unloading success rate. At the same time, when the cluster environment and task changes are complex, the HTO algorithm can still show good stability and can adapt to various complex changes.

[0106] Table 6 Task delay and task offloading success rate in different environments and different task amounts when the task complexity is 50

[0107]

[0108] Table 6 shows the performance of the two algorithms on different indicators when the task complexity is 50, that is, the task is very complex. For the GD algorithm, under the three different task number environment settings of low, medium and high, with the increase of the number of cluster nodes, that is, with the improvement of the overall computing power of the cluster, the average task delay and task offloading failure rate are gradually reduced, which shows that the GD algorithm can effectively utilize the computing resources of the cluster. For the HTO algorithm, its performance is relatively stable in all environments. From the perspective of task delay, HTO can still ensure the stability of cluster task execution by effectively allocating computing resources in a complex environment at the expense of a certain task delay. When the cluster computing resources increase, that is, when the cluster computing resources change, the HTO algorithm can still reasonably allocate computing resources, which proves the advantage of the HTO algorithm in complex environments. In contrast, the GD algorithm cannot effectively adapt and optimize the processing when facing complex environments. From the perspective of task volume, with the increase of task volume, the performance of the two algorithms changes. The decline of the GD algorithm is more obvious, while the HTO algorithm shows better stability and adaptability while ensuring performance.

[0109] The performance of the HTO algorithm is compared vertically. When the cluster computing resources and task volume are consistent, the performance of the HTO algorithm is analyzed as the task complexity increases. As the task complexity increases, the average task delay of the HTO algorithm gradually increases, especially when the computing resources are tight, such as when the task complexity is 40 and 50 and the cluster computing resources are 100 mobile devices and 5 MEC servers. Compared with the task complexity of 20 and 30, the task delay has a small cliff-like increase. The same situation also occurs in the GD algorithm. This is because the task volume obviously exceeds the unit computing power of the cluster. This situation is significantly alleviated when the cluster computing resources are expanded. For example, the mobile200_mec10 setting is particularly obvious compared to the mobile100_mec5 setting when the task complexity is 50. From the perspective of task offloading success rate, compared with the GD algorithm, HTO can obviously sacrifice a certain task delay in exchange for the guarantee of task offloading success rate. The same phenomenon does not occur in the GD algorithm, which shows that in complex environments and insufficient computing resources, HTO can take into account both task delay and task offloading success rate, that is, reduce task delay as much as possible while ensuring successful task offloading. This phenomenon occurs due to the particularity of the reward function setting.

[0110] The above description is only a detailed description of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, according to the ideas provided by the present invention, there will be changes in the specific implementation methods, and these changes should also be regarded as the protection scope of the present invention.

Claims

1. A heterogeneous task allocation method based on predicted balance ratio and resource constraints, characterized in that: The following steps are involved: S1. Obtain the load prediction results of the node load in the mobile edge computing MEC cluster through deep learning algorithms; S2. Based on the load prediction results, the load is migrated through the early warning mechanism of the load migration strategy LMSPER algorithm based on the predicted balancing ratio; S3. Obtain the heterogeneous task queue data of the MEC cluster, make task offloading decisions based on the task queue data and the MEC cluster status, and complete task offloading allocation.

2. The heterogeneous task allocation method based on predicted balance ratio and resource constraints according to claim 1 is characterized in that: The MEC cluster status consists of three parts: the node status uploaded by the nodes in each MEC cluster in real time, the load prediction results of the node load in the EMC cluster, and the past node status and load data.

3. The heterogeneous task allocation method based on predicted balance ratio and resource constraints according to claim 1 is characterized in that: The specific implementation process of the LMSPER algorithm is as follows: S2.

1. Based on the obtained load prediction results, use the load prediction framework to perform load early warning monitoring. When the load early warning monitors that the workload of the node reaches the threshold of 80%, the load migration mechanism is triggered, and the load balancing ratio of each node and the corresponding task is calculated to obtain the load balancing ratio set of the nodes to be processed. The load balancing ratio set contains the load balancing ratio matrix of each node. S2.

2. The load balancing ratio matrix of each node contains the balancing ratio of each task. The balancing ratios of the tasks in each node are sorted from small to large, and the load is migrated in the sorted order until the load prediction result of the node is lower than the load threshold.

4. The heterogeneous task allocation method based on predicted balance ratio and resource constraints according to claim 3 is characterized in that: The calculation of the load balancing ratio of each node to the corresponding task is specifically: Load balancing ratio b ij The calculation expression is: where b ij represents the balance ratio of the i-th task on the node to be migrated to the j-th node, represents the weight of the kth resource of the jth node, represents the expected load of the jth node on the kth resource at a certain moment in the future; the remaining resource matrix W is used to represent the remaining resource status of the remaining nodes in the MEC cluster, and the corresponding mathematical expression is W = (ω1, ω2, …, ω m ), Represents the resource demand vector for the kth resource when the i-th task is migrated to the j-th node.

5. The heterogeneous task allocation method based on predicted balance ratio and resource constraints according to claim 1 is characterized in that: The task offloading decision is implemented through the heterogeneous task offloading HTO algorithm based on resource constraints and meta-reinforcement learning. The specific implementation process is as follows: S4.

1. Obtain tasks from the set of tasks to be processed and record the resource requirements of the tasks; then, the task sampler performs path sampling on the sampled tasks, that is, determines the execution path and unloading method of the tasks according to the resource requirements of the tasks and the current state of the cluster; finally, the task sampling processor is used to format the sampled task data; S4.2, internal strategy update: based on the formatted task data, the proximal strategy optimization PPO algorithm is used to optimize the execution path and offloading method; S4.

3. External strategy update: Through the data and experience accumulated on multiple tasks through meta-reinforcement learning, the experience across multiple tasks is used to improve the learning strategy of task offloading.

Images